Objective
To investigate the effectiveness of a convolutional neural network (CNN) in the detection of healthy teeth and early carious lesions on occlusal surfaces, and to assess the applicability of this deep-learning algorithm as an aid in the diagnosis of dental caries.
Materials and Methods
A total of 2,481 posterior teeth (2,459 permanent and 22 deciduous teeth) with varying stages of carious lesions were classified according to the International Caries Detection and Assessment System (ICDAS). After clinical evaluation, ICDAS 0 and 2 occlusal surfaces were photographed with a professional digital camera. VGG-19 was chosen as the CNN and the findings were compared with those of a reference examiner to evaluate its detection efficiency. To verify the effectiveness of the CNN as a diagnostic aid, three examiners (an undergraduate student (US), a newly graduated dental surgeon (ND), and a specialist in pediatric dentistry (SP) assessed the acquired images (Phase I). In Phase II, the examiners reassessed the same images using the CNN-generated algorithms.
Results
The training dataset consisted of 8,749 images, whereas the test dataset included 140 images. VGG-19 achieved an accuracy of 0.879, sensitivity of 0.827, precision of 0.949, and F1-score of 0.887. In Phase I, the accuracy rates for examiners US, ND, and SP were 0.543, 0.771, and 0.807, respectively. In Phase II, the accuracy rates improved to 0.679, 0.886, and 0.857 for the respective examiners. The number of correct answers was significantly higher in Phase II than in Phase I for all examiners (McNemar test;P < 0.05).
Conclusions
VGG-19 demonstrated satisfactory performance in the detection of early carious lesions and as a diagnostic aid.
Clinical relevance:
Automated detection of early carious lesions by deep-learning algorithms is an important aid in the early diagnosis of the disease, as it minimizes subjective assessments by different examiners, enabling quicker and more reliable clinical decision-making.